Switch from mlx-rs to quill-mlx

crates/voice-cli/Cargo.toml

@@ -24,6 +24,7 @@ pubkey = "RWRmHACtt9SpjQB/l+dAoCe+3iqxq/Xe/90P6dyCW70axzy3UkWl+Ave"
 
 [dependencies]
 voice-tts = { version = "0.2.0", path = "../voice-tts" }
+quill-mlx = "0.25.5"
 voice-g2p = { version = "0.2.0", path = "../voice-g2p" }
 voice-stt = { version = "0.1.0", path = "../voice-stt" }
 clap = { version = "4", features = ["derive"] }

crates/voice-cli/build.rs

@@ -76,7 +76,7 @@ fn find_metallib(build_dir: &Path) -> Option<PathBuf> {
     for entry in entries.flatten() {
         let name = entry.file_name();
         let name_str = name.to_string_lossy();
-        if name_str.starts_with("mlx-sys-") {
+        if name_str.starts_with("quill-mlx-sys-") || name_str.starts_with("mlx-sys-") {
             println!(
                 "cargo:warning=[metallib] found mlx-sys dir: {}",
                 entry.path().display()

crates/voice-cli/src/mcp.rs

@@ -204,6 +204,12 @@ pub fn run(config: ServerConfig) {
         stt_tokenizer: None,
     };
 
+    // Cap the Metal buffer cache at 2 GB to prevent unbounded memory growth
+    // across repeated inference calls.
+    if let Err(e) = quill_mlx::metal::set_cache_limit(2 * 1024 * 1024 * 1024) {
+        eprintln!("warning: failed to set Metal cache limit: {e}");
+    }
+
     let mut stdout = io::stdout();
 
     if !QUIET.load(Ordering::Relaxed) {

crates/voice-dsp/Cargo.toml

@@ -8,4 +8,4 @@ license = "MIT"
 repository = "https://github.com/rgbkrk/voicers"
 
 [dependencies]
-mlx-rs = { version = "0.25" }
+quill-mlx = "0.25.5"

crates/voice-dsp/README.md

@@ -28,7 +28,7 @@ let (magnitude, phase) = stft.transform(&audio_batch)?;
 let reconstructed = stft.inverse(&magnitude, &phase)?;
 ```
 
-All functions operate on `mlx_rs::Array` and return `Result<_, mlx_rs::error::Exception>`.
+All functions operate on `quill_mlx::Array` and return `Result<_, quill_mlx::error::Exception>`.
 
 ## Requirements